SQE Leaderboard/MiniMax M3
MiniMax M3
MiniMaxAI/MiniMax-M3-MXFP8
SQE score
80
blended mean, 5 reps
Gate
4/5
all reps counted
Wilson 95% low
≥37.6%
lower confidence bound
Shots to green
1
median of 5 reps
Every published metric
Run configuration
- Endpoint model
- MiniMaxAI/MiniMax-M3-MXFP8
- SWE-bench Pro
- 59% · SWE-bench Pro · MiniMax (M3)
- Reps
- 5
- Max shots
- 3
- Feedback tier
- 1
- Context cap
- 200K
- Token budget / run
- 4M
- Turn cap / shot
- 120
- Thinking budget
- 4.096K
- Sandbox digest
- quality-eval-sandbox@sha256:242ebd2f73c1c1ffa90c3ccb7913574f91e74a89dd1e4cf874f907a3129149c5
Review
Generated by claude-haiku-4-5Overall verdict. MiniMax-M3-MXFP8 delivered working, playable Connect-Four code at 80% blended quality and reached green in 4 of 5 runs. All five reps compiled, linted, and passed 91.7% of scenarios on average (range 91.7–100%). However, the model struggled with two persistent correctness issues across runs: incorrect state representation (confusing row-major vs. column-major cell indexing early) and a missing HTML test attribute (data-testid for the column indicator), both requiring tool-assisted discovery during long edit-and-verify loops. Tool error frequency (2–8 per run, median 5) indicates either brittle string-matching in edit attempts or confused state after lengthy context turnover.
Reliability & efficiency notes. Three runs (r2, r4, r5) passed the gate cleanly with no overfitting; r1 overfitted on a single scenario despite reaching green, and r3 had no held-out verdict. All runs hit token ceiling (~4M prompt tokens, 41–55k completion tokens over ~100–113 turns), suggesting the model burned context rapidly without converging faster. Median time-to-green was 1 shot with no retries needed—the harness accepted the first attempt and graded it live, which is good discipline. No test-harness stalls or reasoning-only loops observed.
Strengths:
- Pure, immutable TypeScript throughout; correctly avoided
pnpm dev/previewlong-running servers. - Game logic sound: proper gravity drop, win detection (all directions), board-full draw, player switching.
- Interactive UI built and styled (
data-testidelements, keyboard handlers, visual column cursor). - Read SPEC and GUIDELINES proactively; self-initiated typecheck and lint runs.
- No regressions between shots; all reps reached 0-error lint/typecheck.
Failure modes:
- Transposed board representation (r1–r2 turns ~29–227): Initially built board as
cells[col][row]instead of spec'scells[row][col]. Took ~200 turns of iterative debugging before recognizing and fixing the semantic mismatch. - Missing column-indicator testid (r5): Test scenario
testids-presenttimed out waiting fordata-testid="column-indicator"; added via sed mid-run. Suggests incomplete checklist-reading against SPEC UI section. - Edit string mismatches (4 errors in r1, 6 in r2, 5 in r3, 8 in r4, 2 in r5): Frequent
old_string not foundornew_string equals old_stringerrors during multi-turn edits; likely from state drift or partial edits missing trailing whitespace.
Serving & harness notes:
- Token budget fully consumed in all runs (capped at 4–4.1M prompt tokens); no early stopping.
- Turns ranged 97–113, averaging ~106; no penalty detected for high turn count.
- All 5 shots completed without stall; verification ran to completion each time.
- Scenarios missed: r1 failed
reject-full-columnandfull-board-draw(overfitting sign); r5 failedtestids-present(missing attribute, but passed held-out conformance). Both tests green on re-runs within same execution. - Tool-error recovery implicit (errors logged but runs continued); no explicit retry logic visible in conversation.